AI Interview Ethics Checklist: Consent, IRB and DPDP
Ethics checklist for AI-moderated interviews
An AI interviewer doesn’t change the ethics of research. Participants still need to understand what they are agreeing to, ethics committees still need to see your protocol, and data protection law still applies. What changes is where things can go wrong: a consent script nobody tested by phone, an audio file sitting with a subprocessor you never listed, a probe that only follows up with the most articulate people. This checklist is the practical companion to our guide to AI interviews for research.
1. Informed consent, spoken and logged
Participants need to know they are speaking to an AI, who is running the study, why, how long it will take, what is recorded, who will hear it, and that they can stop at any time. On a phone call, consent has to be spoken, in the participant’s language, before any research question is asked, and it has to be logged. J-PAL’s phone-survey advice is to keep the script short, pilot it by phone, and not continue without explicit confirmation. Zolabs is designed to take spoken consent at the start of each interview and end the call if the person declines.
- Disclosure: “Hello, this is an automated AI assistant calling on behalf of [organisation]. I am not a person.”
- Purpose: “We are speaking with people who took part in [programme] to learn [one-line purpose].”
- Time: “The conversation takes about [N] minutes.”
- Recording: “The call is recorded and turned into a written transcript. Only the [organisation] research team and [named processor] will have access to it.”
- Voluntary: “You can skip any question or stop at any time by saying ‘stop’. This will not affect any support you receive from [organisation].”
- Retention and rights: “We keep the recording for [period]. You can ask us to delete it by calling [number].”
- Confirmation: “Do you agree to continue? Please say yes or no.” → No / unclear answer: thank, end the call, log the refusal. Yes: log timestamp, language and consent version, then begin.
Two points from our delivery work. Pilot the script by phone with people like your participants; a sentence that reads well on paper can be confusing when heard once. And make the “this will not affect your support” line explicit for programme participants, who may worry that saying no has consequences.
2. Ethics committee or IRB review
If your study would need ethics review with human interviewers, it needs it with an AI interviewer too. In India that usually means an Institutional Ethics Committee working under the ICMR National Ethical Guidelines for health research, or your university’s committee for social science. In the US and UK it means your IRB or REC. Expect to submit:
- the AI’s full discussion guide and consent script, in every language;
- how distress, refusals and requests to stop are handled;
- which AI and telephony providers process the audio, where, and for how long;
- how humans review AI-generated codes before they become findings.
Committees new to AI interviews also tend to ask what happens if the AI says something off-script. Have an answer: the guide is fixed for the study, prompt changes trigger a fresh pilot, and a named researcher reviews a share of calls in full.
3. DPDP Act and data protection
India’s Digital Personal Data Protection Rules, 2025 were notified in November 2025, with most obligations phased in over 18 months. Voice recordings, phone numbers and transcripts are personal data. Before you call anyone, be clear on:
- Who the data fiduciary is (usually your organisation).
- Who the processors are (the AI interview platform and its subprocessors).
- How long recordings are kept, and when they are deleted.
- How a participant asks for deletion, and who acts on it.
Ask any vendor, including us, for a data processing agreement, a current subprocessor list and a retention schedule in writing. For the notice itself, our DPDP consent notice template for NGOs gives wording you can adapt; for the wider organisational duties, see the DPDP compliance checklist. ZolaShield covers DPDP readiness for nonprofits if you want help.
Children. Under the DPDP Act, anyone under 18 is a child, and processing their data needs verifiable consent from a parent or guardian. We recommend not using AI moderation with children at all; interview caregivers instead. Our guide to verifiable parental consent under DPDP explains the mechanics if caregivers’ data involves their children.
Collect less. Zolabs’ interviewer is instructed not to ask for Aadhaar, PAN, OTPs or card numbers, or about caste. Most studies don’t need any of these, and a recorded call is the worst place to hold them.
4. Bias: who the AI hears, and how it codes
AI moderators can be biased in three places:
- Probing: following up more on articulate answers, so quieter participants give thinner data.
- Coding: NORC found slight over-coding linked to acquiescence (Barari et al., 2026).
- Reach: people with phones, at home, during call hours.
Report results by gender, location and language, and treat the AI’s first-pass themes as a hypothesis, not a finding. A simple check: compare the average answer length and number of probes per interview across subgroups. A large gap tells you where to listen first.
The checklist on one page
- AI disclosure in the first sentence of every call
- Consent script translated, piloted by phone and approved in every language
- Refusal ends the call and is logged
- Stop and distress protocol written, with a named human to follow up
- Ethics committee / IRB approval with guide, script, data flow and review plan
- DPA, subprocessor list and retention schedule from the vendor, in writing
- No identifiers collected that the study doesn’t need
- No AI moderation with under-18s
- Results and coverage reported by gender, location and language
Transcription errors are an ethics issue too, because a mistranscribed “no” becomes a false finding. We cover that in voice AI interviews in Indian languages.
Planning a submission? Bring it to a free 30-minute AI Clinic → We’ll go through your consent script and data flow with you.